OpenAI previewed Ultrafast on August 13 — a new API service tier running GPT-5.6 Sol at 750 output tokens per second on Cerebras wafer-scale chips. The claim: 14 times faster than the standard processing tier. That's not an incremental speed bump. At standard speed, generating a paragraph-length response takes roughly 25 seconds. At 750 TPS, the same output lands in about 1.5 seconds.

The speed difference is architecturally grounded. Cerebras builds wafer-scale chips — the entire silicon wafer is one processor, eliminating the inter-chip communication that makes clustered GPU inference slow on fast-token-generation tasks. For voice AI specifically, first-token latency and sustained throughput are the binding constraints on product quality. This is the problem Ultrafast is designed to solve.

But the product announcement is the less interesting part of this story.

In December 2025, OpenAI and Cerebras signed a Master Relationship Agreement committing OpenAI to purchase 750 megawatts of Cerebras inference compute capacity through 2028, with an option for an additional 1.25 gigawatts — a deal Cerebras values at over $20 billion. OpenAI also provided Cerebras a $1 billion secured loan in January 2026, with warrants attached giving OpenAI the right to purchase Cerebras shares at contractual prices.

In July 2026, OpenAI exercised those warrants for 10,033,508 Class N shares at a contractual exercise price of $0.00001 per share. Total cash outlay: roughly $100. At Cerebras's Class A trading price on August 13 — approximately $229 per share — that holding carried an implied value near $2.3 billion. The shares carry no votes.

Then, a few weeks later, Ultrafast launched.

The sequencing matters. OpenAI committed to Cerebras's infrastructure before there was a product to sell. They locked in equity before the product launched. Ultrafast is, among other things, the product that justifies the infrastructure bet and makes the equity stake worth something in the market narrative. This is not a coincidence. OpenAI didn't just build a fast model tier — they built a reason for Cerebras to exist at frontier scale.

Source spread

• OpenAI — Previewing Ultrafast mode: GPT-5.6 Sol at up to 14× [hype] — first-party announcement; 750 TPS and 14× are vendor-claimed, no independent benchmark cited.
• Cerebras — Accelerating GPT-5.6 Sol Ultrafast with OpenAI [hype] — chip-side technical framing; use-case context (voice AI, incident response, financial research).
• Cerebras Investor Relations — Ultrafast press release [builder] — source for warrant exercise details: share count, exercise price, implied value, Class N structure.
• CNBC — Cerebras scores OpenAI deal worth over $10 billion [builder] — original January 2026 reporting on the $1B loan and the 750MW compute commitment.
• Implicator — OpenAI paid $100 for a 4.2% Cerebras stake [skeptic] — the analysis that surfaced the warrant-exercise timing and what the sequence implies.

Pros & cons

What's real:

• The speed difference is architecturally grounded, not marketing. Cerebras wafer-scale chips genuinely eliminate inter-chip latency that makes multi-GPU clusters slow at high-throughput, short-context generation. The 750 TPS claim is vendor-sourced, but it has a real mechanism behind it.
• Real-time voice AI has been latency-constrained since day one. First-token latency and sustained output speed are the binding constraints on voice product quality, and 14× faster directly addresses both. If the numbers hold in production, Ultrafast unblocks a category of voice product that wasn't buildable at frontier model quality.
• OpenAI holding a 4.2% Cerebras stake aligns incentives between OpenAI and a key infrastructure dependency. The risk of Cerebras deprioritizing the OpenAI relationship is lower than it would be with a pure vendor arrangement.

What deserves a side-eye:

• No published price and no GA date. "Limited preview" is the hedging language you use when you're not confident how the infrastructure holds under broad load. Don't build a product on a tier that has no price page.
• The 750 TPS figure is vendor-claimed and uncorroborated by independent benchmarks. Wait for Artificial Analysis or similar third-party evaluators before committing architecture decisions to it.
• 750 megawatts of committed compute is a large infrastructure bet. If Ultrafast adoption lags, the economics of that commitment look very different — and OpenAI is partially incentivized to push adoption through product launches exactly like this one.

Samwise's take

What builders need to know

Further reading

• OpenAI — Previewing Ultrafast mode: GPT-5.6 Sol at up to 14× the speed — official announcement and preview access link
• Cerebras — Accelerating GPT-5.6 Sol Ultrafast with OpenAI — chip-side technical framing and use-case context
• Cerebras Investor Relations — Cerebras Powers Ultrafast Mode for OpenAI's GPT-5.6 Sol — warrant exercise details, equity structure
• CNBC — Cerebras scores OpenAI deal worth over $10 billion — original January 2026 reporting on the $1B loan and compute commitment
• Implicator — OpenAI paid $100 for a 4.2% Cerebras stake weeks before Ultrafast launch — warrant-exercise timing analysis